AI-Enhanced Anomaly Detection with ARIMA Rainfall Forecasting

Abstract

Finding anomalous patterns in large datasets that may indicate security risks or system malfunctions is a crucial task in cloud computing. Recent advances in machine learning, combined with the rapid growth of cloud computing, have enabled several cloud-based, data-driven approaches for autonomous anomaly detection. However, conventional anomaly detection methods are often resource- and time-intensive, particularly when processing massive volumes of data. The method proposed in this study addresses these challenges by providing a scalable and efficient mechanism for real-time identification of irregularities. Furthermore, the approach supports timely detection and resolution of system malfunctions, ensuring the continuous availability and effectiveness of cloud-based systems. The investigation utilizes a 50 GB NCEP (National Centers for Environmental Prediction) time-series dataset covering the period from 2010 to 2024 with a daily temporal resolution. The selected geographic region spans 77° longitude by 12° latitude, focusing on the Bangalore region. Since the Z-score is a key statistical measure for anomaly identification, it is computed to detect deviations from normal behavior. Monthly anomaly detection is performed for precipitation values from 2010 to 2024 to analyze temporal variations. In addition, a comparative analysis is carried out between ARIMA, SARIMA, and existing machine learning models to evaluate performance, accuracy, scalability, and robustness.